double CostFromErrFunc::value(const vector<double>& xin, Model* model) {
VectorXd x = getVec(xin, vars_);
VectorXd err = f_->call(x);
if (coeffs_.size()>0) err.array() *= coeffs_.array();
switch (pen_type_) {
case SQUARED: return err.array().square().sum();
case ABS: return err.array().abs().sum();
case HINGE: return err.cwiseMax(VectorXd::Zero(err.size())).sum();
default: assert(0 && "unreachable");
}
return 0; // avoid compiler warning
}
void drwnNNGraphMLearner::subGradientStep(const MatrixXd& G, double alpha)
{
DRWN_FCN_TIC;
// gradient step
_M -= alpha * G;
// project onto psd
SelfAdjointEigenSolver<MatrixXd> es;
es.compute(_M);
const VectorXd d = es.eigenvalues().real();
if ((d.array() < 0.0).any()) {
const MatrixXd V = es.eigenvectors();
_M = V * d.cwiseMax(VectorXd::Constant(d.rows(), DRWN_EPSILON)).asDiagonal() * V.inverse();
}
DRWN_FCN_TOC;
}
